com.sky.sentis.yolox

sentis-yolox

YOLOX object detection converted to Unity Inference Engine (Sentis) FP16.

Files

yolox_fp16.sentis
YoloxDetector.cs    # self-contained Unity Sentis inference (COCO labels)

Model I/O

• Input: RGB NCHW [1, 3, S, S] in [0, 1] (read S from the model; the [0,1]→[0,255] scale is baked into the export). YOLOX is typically trained BGR, so a R↔B swizzle is usually needed.
• Output: one tensor [1, N, 5+C] (or [1, 5+C, N]) where each candidate is [cx, cy, w, h, objectness, class_0..class_{C-1}], scores already sigmoid-activated and boxes in input-pixel coordinates (export with the grid/stride decode baked in). NMS runs on the C# side.

Inference

A complete self-contained implementation lives in YoloxDetector.cs (texture → tensor with optional R↔B swizzle, layout auto-detection, scoring, class-wise NMS, GPU warmup, 80 COCO labels inlined). Minimal usage:

var detector = new YoloxDetector(BackendType.GPUCompute);
detector.Load(modelRoot);                                  // folder holding yolox_fp16.sentis
List<YoloxDetector.Detection> dets = await detector.DetectAsync(frameTexture); // boxes in input-pixel space

Do not substitute Ultralytics weights (yolo*), which are AGPL-3.0. This repo ships YOLOX only.

License & attribution

Apache-2.0. Converted from Megvii-BaseDetection/YOLOX (Apache-2.0).